@agentmem/mastra

Mastra integration for AgentMem — drop-in memory tools for Mastra agents.

npm install @agentmem/mastra @mastra/core zod

Quick start

import { Agent } from '@mastra/core/agent'
import {
  AgentMemIntegration,
  agentmemMemorize,
  agentmemRemember,
} from '@agentmem/mastra'

const agentmem = new AgentMemIntegration({
  apiKey:  process.env.AGENTMEM_API_KEY!,
  agentId: 'support-bot',            // default; tools can override per-call
  scope:   'team',                   // optional default
})

const agent = new Agent({
  name: 'support',
  instructions:
    'You help customers. When you learn a durable fact (preference, decision, ' +
    'constraint), call `memorize`. Before answering, call `remember` to recall ' +
    'prior context.',
  tools: {
    memorize: agentmemMemorize(agentmem),
    remember: agentmemRemember(agentmem),
  },
})

What it gives you

| Export | Purpose | |---|---| | AgentMemIntegration | Class wrapping the AgentMem SDK; exposes createMemory / searchMemory / searchMemoryRaw. | | agentmemMemorize(integration, defaults?) | Pre-built Mastra tool. Stores a fact. | | agentmemRemember(integration, defaults?) | Pre-built Mastra tool. Retrieves relevant memories. | | MemoryStore, AgentMemError, Scope, Role, MemoryHit | Re-exported from @agentmem/sdk for convenience. |

Defaults and overrides

Constructor AgentMemConfig accepts default agentId, scope, role, workflowId. Every per-call site (tool input or method argument) can override. The merge order is:

tool input > tool factory defaults > integration config > error if still missing.

agentId is the only required field; if it isn't set on the integration AND isn't passed per-call, the tool throws a helpful error.

Using the integration without tools

If you don't want the pre-built tools (e.g., you're writing custom flows):

const text = await agentmem.searchMemory('what does the customer prefer?', {
  agentId: 'support-bot',
  topK:    5,
})
// → prompt-prefixed string ready to inject into the model's context

const ack = await agentmem.createMemory('Customer prefers email over phone.', {
  agentId: 'support-bot',
  scope:   'team',
})
// → "Memory saved (id: ...). Content: ..."

For raw results without the prompt prefix:

const hits = await agentmem.searchMemoryRaw('escalation threshold', { agentId: 'support-bot' })
// → MemoryHit[]

License

Apache-2.0